Neural network approach for parameter learning to speed up planning for complex driving scenarios

Abstract

In one embodiment, a computer-implemented method of operating an autonomous driving vehicle (ADV) includes perceiving a driving environment surrounding the ADV based on sensor data obtained from one or more sensors mounted on the ADV, determining a driving scenario, in response to a driving decision based on the driving environment, applying a predetermined machine-learning model to data representing the driving environment and the driving scenario to generate a set of one or more driving parameters, and planning a trajectory to navigate the ADV using the set of the driving parameters according to the driving scenario through the driving environment.

What is claimed is: 
     
       1. A computer-implemented method of operating an autonomous driving vehicle (ADV), comprising:
 perceiving, by a processor, a driving environment surrounding the ADV based on sensor data obtained from one or more sensors mounted on the ADV; 
 determining, by the processor, a driving scenario, in response to a driving decision based on the driving environment; 
 applying, by the processor, a predetermined machine-learning model to data representing the driving environment and the driving scenario to generate a set of driving parameters, wherein the set of driving parameters include a speed, an acceleration, an acceleration penalty, a curvature, a curvature penalty, and a heading direction, and wherein the acceleration penalty and the curvature penalty are coefficients of a cost function to determine an optimal acceleration penalty and an optimal curvature penalty; 
 determining the cost function=(P a ×α)+(P k ×κ), wherein Pa is the acceleration penalty, α (alpha) is the acceleration, Pk is the curvature penalty, κ (kappa) is the curvature, and Pa and Pk are coefficients of the cost function, and wherein when the driving scenario is a U-turn driving scenario, the optimal Pa is equal to 0.7, and the optimal Pk is equal to 0.3; 
 planning, by the processor, a trajectory to navigate the ADV using the set of the driving parameters according to the driving scenario through the driving environment based on the cost function; and 
 controlling, by the processor, the ADV to navigate according to the trajectory. 
 
     
     
       2. The method of  claim 1 , further comprising:
 extracting a set of features from one or more images representing the perceived driving environment; and 
 providing the set of features to one or more inputs of the predetermined machine-learning model as a part of the data representing the driving environment. 
 
     
     
       3. The method of  claim 2 , further comprising:
 determining a reference line of a road in which the ADV is moving based on the one or images and a map associated with the road; and 
 providing the reference line to the one or more inputs of the predetermined machine-learning model as a part of the data representing the driving environment. 
 
     
     
       4. The method of  claim 1 , wherein planning a trajectory to navigate the ADV using the set of the driving parameters comprises:
 generating a set of trajectory candidates based on the driving scenario in view of the driving environment; and 
 for each of the trajectory candidates, optimizing the trajectory candidate using the set of driving parameters to derive an optimized trajectory to drive the ADV. 
 
     
     
       5. The method of  claim 1 , wherein optimizing a trajectory candidate using the set of driving parameters comprises:
 determining the cost function, including adjusting one or more coefficients of the cost function based on the set of driving parameters; 
 calculating a cost for each of the trajectory candidates using the cost function; and 
 selecting one of the trajectory candidates having a lowest cost as the optimized trajectory. 
 
     
     
       6. The method of  claim 1 , wherein the driving scenario comprises at least one of the U-turn scenario, a turn-left scenario, a turn-right scenario, a moving straight scenario, a lane changing scenario, or a parking scenario. 
     
     
       7. The method of  claim 1 , wherein the machine-learning model was trained based on driving statistics collected from a plurality of vehicles driving under a plurality of driving scenarios. 
     
     
       8. A non-transitory machine-readable medium having instructions stored therein, which when executed by a processor, cause the processor to perform operations of operating an autonomous driving vehicle (ADV), the operations comprising:
 perceiving a driving environment surrounding the ADV based on sensor data obtained from one or more sensors mounted on the ADV; 
 determining a driving scenario, in response to a driving decision based on the driving environment; 
 applying a predetermined machine-learning model to data representing the driving environment and the driving scenario to generate a set of driving parameters, wherein the set of driving parameters include a speed, an acceleration, an acceleration penalty, a curvature, a curvature penalty, and a heading direction, and wherein the acceleration penalty and the curvature penalty are coefficients of a cost function to determine an optimal acceleration penalty and an optimal curvature penalty; 
 determining the cost function=(P a ×α)+(P k ×κ), wherein Pa is the acceleration penalty, α (alpha) is the acceleration, Pk is the curvature penalty, κ (kappa) is the curvature, and Pa and Pk are coefficients of the cost function, and wherein when the driving scenario is a U-turn driving scenario, the optimal Pa is equal to 0.7, and the optimal Pk is equal to 0.3; 
 planning a trajectory to navigate the ADV using the set of the driving parameters according to the driving scenario through the driving environment based on the cost function; and 
 controlling, by the processor, the ADV to navigate according to the trajectory. 
 
     
     
       9. The non-transitory machine-readable medium of  claim 8 , wherein the operations further comprise:
 extracting a set of features from one or more images representing the perceived driving environment; and 
 providing the set of features to one or more inputs of the predetermined machine-learning model as a part of the data representing the driving environment. 
 
     
     
       10. The non-transitory machine-readable medium of  claim 9 , wherein the operations further comprise:
 determining a reference line of a road in which the ADV is moving based on the one or images and a map associated with the road; and 
 providing the reference line to the one or more inputs of the predetermined machine-learning model as a part of the data representing the driving environment. 
 
     
     
       11. The non-transitory machine-readable medium of  claim 8 , wherein planning a trajectory to navigate the ADV using the set of the driving parameters comprises:
 generating a set of trajectory candidates based on the driving scenario in view of the driving environment; and 
 for each of the trajectory candidates, optimizing the trajectory candidate using the set of driving parameters to derive an optimized trajectory to drive the ADV. 
 
     
     
       12. The non-transitory machine-readable medium of  claim 8 , wherein optimizing a trajectory candidate using the set of driving parameters comprises:
 determining the cost function, including adjusting one or more coefficients of the cost function based on the set of driving parameters; 
 calculating a cost for each of the trajectory candidates using the cost function; and 
 selecting one of the trajectory candidates having a lowest cost as the optimized trajectory. 
 
     
     
       13. The non-transitory machine-readable medium of  claim 8 , wherein the driving scenario comprises at least one of the U-turn scenario, a turn-left scenario, a turn-right scenario, a moving straight scenario, a lane changing scenario, or a parking scenario. 
     
     
       14. The non-transitory machine-readable medium of  claim 8 , wherein the machine-learning model was trained based on driving statistics collected from a plurality of vehicles driving under a plurality of driving scenarios. 
     
     
       15. A data processing system for operating an autonomous driving vehicle (ADV), comprising:
 a processor; and 
 a memory coupled to the processor to store instructions, which when executed by the processor, cause the processor to perform operations, the operations including:
 perceiving a driving environment surrounding the ADV based on sensor data obtained from one or more sensors mounted on the ADV, 
 determining a driving scenario, in response to a driving decision based on the driving environment, 
 applying a predetermined machine-learning model to data representing the driving environment and the driving scenario to generate a set of driving parameters, wherein the set of driving parameters include a speed, an acceleration, an acceleration penalty, a curvature, a curvature penalty, and a heading direction, and wherein the acceleration penalty and the curvature penalty are coefficients of a cost function to determine an optimal acceleration penalty and an optimal curvature penalty, 
 determining the cost function=(P a ×α)+(P k ×κ), wherein Pa is the acceleration penalty, α (alpha) is the acceleration, Pk is the curvature penalty, κ (kappa) is the curvature, and Pa and Pk are coefficients of the cost function, and wherein when the driving scenario is a U-turn driving scenario, the optimal Pa is equal to 0.7, and the optimal Pk is equal to 0.3; 
 planning a trajectory to navigate the ADV using the set of the driving parameters according to the driving scenario through the driving environment based on the cost function, and 
 controlling, by the processor, the ADV to navigate according to the trajectory. 
 
 
     
     
       16. The data processing system of  claim 15 , wherein the operations further comprise:
 extracting a set of features from one or more images representing the perceived driving environment; and 
 providing the set of features to one or more inputs of the predetermined machine-learning model as a part of the data representing the driving environment. 
 
     
     
       17. The data processing system of  claim 16 , wherein the operations further comprise:
 determining a reference line of a road in which the ADV is moving based on the one or images and a map associated with the road; and 
 providing the reference line to the one or more inputs of the predetermined machine-learning model as a part of the data representing the driving environment. 
 
     
     
       18. The data processing system of  claim 15 , wherein planning a trajectory to navigate the ADV using the set of the driving parameters comprises:
 generating a set of trajectory candidates based on the driving scenario in view of the driving environment; and 
 for each of the trajectory candidates, optimizing the trajectory candidate using the set of driving parameters to derive an optimized trajectory to drive the ADV. 
 
     
     
       19. The data processing system of  claim 15 , wherein optimizing a trajectory candidate using the set of driving parameters comprises:
 determining the cost function, including adjusting one or more coefficients of the cost function based on the set of driving parameters; 
 calculating a cost for each of the trajectory candidates using the cost function; and 
 selecting one of the trajectory candidates having a lowest cost as the optimized trajectory. 
 
     
     
       20. The data processing system of  claim 15 , wherein the driving scenario comprises at least one of the U-turn scenario, a turn-left scenario, a turn-right scenario, a moving straight scenario, a lane changing scenario, or a parking scenario. 
     
     
       21. The data processing system of  claim 15 , wherein the machine-learning model was trained based on driving statistics collected from a plurality of vehicles driving under a plurality of driving scenarios.